Bayesian CSP and Motor Imagery: Building a Robust EEG Decoder with NimbusLDA

Imagine a user imagining a left-hand grip. Somewhere across their scalp, a distinct spatial pattern of mu and beta oscillations flickers into existence. Your pipeline captures it, filters it, and fires a command — left. But how sure are you? Ninety percent? Sixty? The same label comes out either way.

That's the quiet problem with the most common motor imagery (MI) BCI pipeline: Common Spatial Patterns (CSP) followed by classical Linear Discriminant Analysis (LDA) is fast, interpretable, and well-validated — but it treats every prediction as equally confident. For lab benchmarks, this is fine. For any system you intend to deploy on a real user, it's a liability.

This post walks through upgrading that pipeline to a fully Bayesian one using NimbusLDA, and explains why calibrated uncertainty is the missing ingredient in most MI decoders.

Why Classical CSP+LDA Falls Short

CSP is a supervised spatial filtering algorithm. It learns a set of filters that maximise the variance ratio between two classes — say, left-hand and right-hand imagery. The filtered signals are log-variance features that feed into a downstream classifier.

Classical LDA then draws a linear boundary in that feature space. It works well, but has two significant limitations for production BCI:

1. No calibrated uncertainty output. LDA produces a class label, and optionally a raw decision-function score — but that score is not a probability. You cannot directly interpret it as confidence.
2. Point-estimate parameters. The class means and shared covariance are estimated once from training data. If the test distribution drifts — as EEG reliably does across sessions and within long recordings — there is no mechanism to express growing uncertainty in the model's own estimates.

In a controlled lab setting, these limitations are easy to paper over. In a clinical trial, an assistive device, or any system where a spurious command has a real cost, they become design flaws.

What Bayesian LDA Actually Does

Bayesian LDA replaces the point-estimate parameters of classical LDA with full posterior distributions over the class means and shared covariance matrix. Instead of asking what is the most likely mean for the left-hand class?, it asks what is the distribution over possible means, given the training data?

This propagates uncertainty all the way through to predictions. For a given test feature vector, the model returns a posterior predictive distribution over classes — a proper probability, not a score.

Three things follow directly from this:

• Calibrated confidence scores. P(left | features) = 0.91 means roughly 91% of trials with that score should be correctly classified. You can verify this empirically with a reliability diagram.
• Principled regularisation. The Bayesian prior over class parameters acts as regularisation. Sparse training data — common in clinical populations where session time is limited — degrades classical LDA faster than Bayesian LDA, because the prior prevents extreme parameter estimates from small samples.
• Honest uncertainty under covariate shift. When a test trial's features fall in a region not well covered by training data, the posterior predictive variance increases. The model expresses that it doesn't know — and you can act on that signal.

Building the Pipeline in Nimbus Studio

Nimbus Studio makes this pipeline visual, parameterised, and reproducible. Here is the node structure:

1. Data node — connect to a MOABB benchmark dataset (e.g., BNCI2014-001, the standard 4-class MI dataset) or stream directly from hardware via BrainFlow.
2. Bandpass filter — 8–30 Hz to isolate the mu and beta rhythms relevant to motor imagery.
3. Epoching — extract 0.5–2.5 s windows following each imagery cue, baseline-corrected.
4. CSP node — fit spatial filters on training epochs. Start with 3 filter pairs (6 total) for binary MI; Nimbus exposes this as a configurable integer parameter.
5. Log-variance features — the standard transform of CSP-filtered epoch data into a compact feature vector.
6. NimbusLDA node — a drop-in replacement for classical LDA. Configure the prior precision (a single scalar hyperparameter controlling regularisation strength) and enable confidence score output.

Once assembled, Nimbus Studio's sweep interface lets you test multiple combinations — different bandpass ranges, filter counts, prior strengths — across the full dataset in minutes, rather than the hours of scripting that would otherwise be required.

Reading and Using Confidence Scores

NimbusLDA outputs two values per trial: a predicted class and a confidence in [0, 1]. A well-calibrated model should produce scores that track empirical accuracy — if the model reports 0.90 confidence on 100 trials, about 90 of those should be correct.

You can verify calibration in Nimbus Studio's evaluation panel using a reliability diagram. If the model is systematically overconfident, reduce the prior precision. If it is underconfident, increase it — or check whether your CSP filters are noisy due to insufficient training data.

In online deployment, a practical three-zone decision strategy works well:

• High confidence (> 0.85): issue the command immediately.
• Medium confidence (0.60–0.85): hold and accumulate evidence over the next trial by multiplying likelihoods.
• Low confidence (< 0.60): discard and prompt the user to retry.

The evidence-accumulation approach is a natural fit for Bayesian inference — you are simply updating the posterior across successive trials until it crosses a decision threshold. This is a lightweight form of Active Inference: the system acts only when its internal model is sufficiently certain, and continues gathering information otherwise.

From Offline Validation to Live Inference

One of the persistent friction points in BCI development is the gap between the offline training pipeline and the online inference system. Preprocessing parameters, spatial filters, and classifier weights typically have to be serialised by hand and re-implemented in a separate real-time codebase. It is the source of a significant fraction of the reproducibility problems that plague published BCI results.

In Nimbus Studio, the same pipeline node graph runs in both modes. Once you click Deploy, the trained CSP filters and NimbusLDA posterior are applied to incoming EEG epochs in real time. Live confidence scores appear in the output panel. No code rewrite is required — the training-to-deployment gap is zero.

For teams using the NimbusSDK directly, the trained model state can be exported and loaded into a Python or Julia inference loop with a few lines of code, maintaining the same Bayesian posterior computed during training.

Conclusion

CSP+LDA has earned its status as the canonical MI BCI pipeline. But classical LDA was designed for statistical discrimination, not for production systems that need to know when to act and when to wait. Bayesian LDA — as implemented in NimbusLDA — is a drop-in upgrade that adds calibrated confidence, principled regularisation, and honest uncertainty quantification without sacrificing the speed or interpretability that make CSP-based pipelines attractive in the first place.

If you are building a motor imagery decoder and you have accepted unlabelled point predictions as the default output, it is worth reconsidering that default. Confidence scores are not a feature — they are the mechanism that makes a BCI safe enough to deploy outside the lab.